DAB-Sky Multitemporal Drone Dataset, 2025
Description
The DAB-Sky dataset comprises 39,827 multitemporal aerial images that integrate both real and synthetic drone imagery for use in drone detection, segmentation, and airspace monitoring research. Real-world images were collected using DSLR and 2K industrial surveillance cameras positioned at varying distances and altitudes. To capture illumination diversity, image acquisition was conducted during four distinct time periods: morning, noon, evening, and night. Complementing the real imagery, synthetic images were produced using Blender 3D and Unreal Engine, where drone models were rendered under adjustable lighting conditions, sky environments, and atmospheric scattering. Aerodynamic variations—including pitch shifts of ±25°, roll adjustments of ±20°, and yaw movements of ±30° were incorporated to simulate realistic flight dynamics, alongside additional visual effects such as motion blur, exposure variation, and HDRI-based illumination. Every image in the dataset is accompanied by two forms of annotations: YOLO-format bounding boxes for object detection tasks and polygon segmentation masks for fine-grained contour analysis. All images were originally recorded at a resolution of 1920×1080 pixels and subsequently standardized to 450×450 pixels to ensure uniformity across the dataset. Through its combination of diverse lighting conditions, varied backgrounds, multiple drone orientations, and dual-domain (real and synthetic) representation, DAB-Sky provides a comprehensive resource suitable for benchmarking drone detection methods, training computer vision models, and supporting research in intelligent airspace surveillance.
Files
Steps to reproduce
The DAB-Sky dataset was created through a multi-stage process that integrates real image acquisition, synthetic rendering, annotation, and systematic dataset partitioning. Real drone images were captured using DSLR and 2K security cameras at varying distances and altitudes, with recordings conducted during morning, noon, evening, and night to achieve illumination diversity. Synthetic images were generated in Blender 3D and Unreal Engine through the modeling of drone geometries and the simulation of aerodynamic motion, atmospheric scattering, motion blur, and HDRI-based lighting environments. All images were annotated using Roboflow, yielding COCO-format JSON files that include bounding-box annotations, polygon segmentation masks, and category labels. Following annotation, the dataset was divided into training, validation, and testing subsets using a 70–20–10 distribution. Reproducing the dataset construction process may require Blender or Unreal Engine for synthetic rendering, Roboflow for annotation, Python for dataset structuring, and standard computer vision libraries capable of reading COCO-style JSON annotations. Adequate storage capacity and GPU resources are recommended for large-scale processing.
Institutions
- Universitas Ahmad Dahlan